engleski

Machine learning guided genetic algorithm for the discovery of novel antimicrobial peptides

By exploring chemical space, researchers try to find novel compounds with favourable features, such as anticancer, antimicrobial or antiviral activity, to combat antibiotic resistant bacteria, facilitate drug delivery or discover new therapeutics. With in vitro experiments being time- and resource-intensive, interest in computationally assisted exploration of chemical space is on the rise. In silico methods can quickly screen thousands of compounds in a matter of hours, filter the most prosperous ones, and thereby speed-up the process while saving resources. In this paper, we present a genetic algorithm guided by machine learning model for the discovery of novel antimicrobial peptides. Firstly, we train a random forest model to differentiate between antimicrobial and non-antimicrobial peptides. The model achieved an accuracy of 88.9%, an F1 score of 87.6%, and an AUC of 88.8%, and was used as a fitness functions the genetic algorithm tries to maximize, which guides it towards novel compounds. Finally, we show that, as the algorithm progresses, the percentage of peptides with high antimicrobial predisposition in population rises from 0% to 100% in 34 iterations. Newly discovered peptides, such as ITIVPKKCKLLL, are then additionally checked by CAMPR3 artificial intelligence antimicrobial peptides prediction tool. Since peptide design is NP-hard, this presents a leap in our endeavours to facilitate in silico discovery of novel valuable compounds.

Machine learning ; Genetic algorithm ; Antimicrobial ; Peptides

nije evidentirano